Hierarchical-based fast heterogeneous information network representation learning method and device

Abstract

The present invention provides a hierarchically structured, fast heterogeneous information network representation learning method and device, comprising first acquiring the data to be mined, performing data cleaning, processing, and encoding, and constructing a heterogeneous information network; then, based on the heterogeneity and topological structure of the network, iteratively reducing the network using a heterogeneous reduction network model to obtain a series of hierarchical heterogeneous information networks of decreasing scale; then, learning the node representation of the smallest heterogeneous information network, and using this node representation to train a heterogeneous refinement model; finally, iteratively refining the network based on the mapping relationship between the heterogeneous refinement model and the hierarchical network to obtain the node representation of the constructed heterogeneous information network. The present invention can reduce the scale of large-scale heterogeneous information networks through reduction technology, thereby reducing the time and space complexity of the representation problem and rapidly learning the node representation in the network.

Description

Technical Field

[0001] The present invention relates to the technical field of heterogeneous information network representation learning, and in particular to a hierarchical structure-based fast heterogeneous information network representation learning method and device. Background Art

[0002] With the advent of the big data era, big data contains a wide variety of objects and their complex relationships. Abstracting these complex interactions using ordinary networks is difficult, but heterogeneous information networks, with their diverse types of nodes and edges, can more fully and naturally analyze such complex data. Heterogeneous information network representation learning, a key tool for mining heterogeneous information networks, aims to embed the rich semantic information and complex structural information in the network into a low-dimensional vector space. These vectors can effectively serve downstream tasks such as node classification, network visualization, and community discovery, thereby better exploring the universal patterns existing in the network. Therefore, heterogeneous information network representation learning has important research value.

[0003] Currently, representation learning methods for heterogeneous information networks can be divided into two main categories: shallow models and deep models. Shallow models primarily preserve local structural information within the network by randomly sampling meta-paths, meta-graphs, and network patterns at a single level. Deep models utilize deep neural networks to learn node representations from interactions between local neighboring nodes within the current network level. These existing methods primarily focus on local similarities between nodes in the network, while ignoring the hierarchical structure present in heterogeneous information networks. Furthermore, the time complexity of representation learning for heterogeneous information networks is proportional to the size of the network. As the network size increases, the time cost required becomes significant. Summary of the Invention

[0004] The present invention proposes a fast heterogeneous information network representation learning method based on hierarchical structure, which is used to retain the hierarchical structure in the heterogeneous information network during the representation learning process to obtain higher quality node representation and accelerate the speed of representation learning.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The hierarchical structure-based fast heterogeneous information network representation learning method proposed in the present invention includes the following steps:

[0007] S1. Obtain the data to be mined, clean and process the data, encode the data, and build a heterogeneous information network ;

[0008] S2, network-based The heterogeneity and topological structure in the network can reduce the network size to obtain a small-scale heterogeneous information network. as well as and Node correspondence between The process of iteratively shrinking the network obtains a series of hierarchical heterogeneous information networks with decreasing network size. and the set of mapping matrices between network layers ,like Representation Network and The mapping matrix between Represents the maximum number of layers in a hierarchical heterogeneous information network;

[0009] S3. Use representation learning algorithm to obtain the smallest Layered Heterogeneous Information Network Node representation . Design a heterogeneous graph convolutional model and based on Train the model to obtain model parameters;

[0010] S4. Using the mapping matrix The network Node representation Pass back to the network ,get The initialization vector of the node is then obtained through the heterogeneous graph convolution model trained in S3 to obtain the final representation of the node in this layer, and the above process is repeated until the original heterogeneous information network , and finally obtain the node representation of the constructed heterogeneous information network.

[0011] As a preferred solution of the above technology, step S1 specifically includes:

[0012] S11, take the academic paper network as an example; first extract the paper title, author and citation from the academic paper database; number them in sequence, such as the paper number from 0 to , author number from arrive , citation number from arrive ; Divide the papers into several categories according to their research fields and number them (e.g. ), and set the category number as the label of the paper.

[0013] S12. Set papers, authors, and citations as different types of nodes, set paper-author edges based on the relationship between authors publishing papers, and set paper-citation edges based on the relationship between papers and citations. Construct a heterogeneous information network based on the above different types of nodes and the relationships between nodes. ,in and Represent the node set and edge set in the network respectively, Represents a set of node and edge types, and Represent the node type mapping function and edge type mapping function respectively. For each node There is a node type mapping function For each edge There is an edge type mapping function ,in and represent the types of nodes and edges in the network, respectively, and and .

[0014] As a preferred solution of the above technology, step S2 specifically includes:

[0015] S21. Design a model to reduce the size of heterogeneous information networks , iterative use Reduce the network scale and obtain a series of hierarchical heterogeneous information networks with decreasing network scales For any layer of heterogeneous information network ,use Based on network heterogeneity and network topology, heterogeneous information network Small-scale heterogeneous information networks formed by reducing scale ,Right now .Model The first step is to divide the heterogeneous information network into The division result is A collection of societies ,gather Any community in Contains multiple types of nodes.

[0016] S22, Model The second step is to pass the collection Any community in China Decompose into multiple type communities according to node type, represented as ,in, Indicates the Layer 1 heterogeneous information network In the community Type community of type, a type community means that all nodes in the community are of the same type. All communities in the network are rearranged according to the node type to obtain a heterogeneous information network. The results of community division by type ,in The type community set is represented as , that is, the node types contained in the community are Type, where express There are indivual Type community.

[0017] S23. Based on heterogeneous information network The results of community division by type , reduce the size of the network to obtain the network .Will The type community in The nodes in the type community are aggregated into The edges between the midpoints. Based on the above idea, we can get and The mapping matrix between , The number of rows is The number of nodes in The number of columns is The number of nodes, that is The number of medium-type communities. The value of is {0,1}, for the type community All nodes in the matrix The corresponding rows in The corresponding column coordinates are treated as 1. The third step is based on and The mapping matrix between Zoom out get The specific formula is as follows:

[0018]

[0019] in, Representing heterogeneous information networks The adjacency matrix of Represents a heterogeneous information network after downscaling The adjacency matrix of .

[0020] S24, Model-based In the construction of heterogeneous information network Iteratively shrink the network to obtain a series of hierarchical heterogeneous information networks with scales from large to small .

[0021] As a preferred solution of the above technology, step S3 specifically includes:

[0022] S31. In the smallest heterogeneous information network The existing unsupervised heterogeneous information network representation learning method is used to learn the representation of nodes ,in , Represents the dimension of the low-dimensional vector.

[0023] S32. Design a heterogeneous graph convolutional model, called a heterogeneous refinement model , first use the Layered Heterogeneous Information Network Node representation To train a heterogeneous refinement model , the first step is to process In order to solve the heterogeneity in the model, different types of node representations are embedded into the same space through the type parameter matrix, and then the Layer convolutional neural network is used to propagate information. No. Taking the layer neural network as an example, is a node In the neural network The hidden layer state in the layer, the specific formula is as follows:

[0024]

[0025] in, , It is The trainable weight matrix of the layer neural network, Indicates the The dimension of the layer representation, Indicates the The node type in the layer neural network is The corresponding type parameter matrix, Representation node of The training optimization function formula is as follows:

[0026]

[0027] in, Representing heterogeneous information networks The adjacency matrix of express The node vector of the final output of the layer convolutional neural network.

[0028] As a preferred solution of the above technology, step S4 specifically includes:

[0029] S41, heterogeneous refinement model trained based on S3 , from the smallest heterogeneous information network A heterogeneous information network will be constructed by layer-by-layer refinement First, use the mapping relationship between different layers of the network The embedding vector of the small-scale network is mapped to the larger-scale network as the initialization node vector of the larger-scale network. Refinement For example, the specific formula is as follows:

[0030]

[0031] S42. Obtaining a heterogeneous information network Initialization vector Then, using the heterogeneous refinement model It realizes the retention of structural information within the same level and the transmission of information between levels during the refinement process.

[0032] S43, based on the two steps S41 and S42, from the smallest network Refine the heterogeneous information network constructed , get the final node representation .

[0033] From the above technical solution, it can be seen that the hierarchical structure-based fast heterogeneous information network representation learning method of the present invention has the following beneficial effects:

[0034] The hierarchical structure-based fast heterogeneous information network representation learning method can reduce the scale of large-scale heterogeneous information networks through scale reduction technology, thereby reducing the time and space complexity of the representation problem to quickly learn the node representation in the network. In addition, this technical solution fully considers the heterogeneity in the network during the reduction and refinement process, and integrates the rich semantic information in the heterogeneous information network into the node representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of a hierarchical structure-based fast heterogeneous information network representation learning method provided by an embodiment of the present invention;

[0036] Figure 2 A framework diagram of a hierarchical structure-based fast heterogeneous information network representation learning method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0038] like Figure 1 and Figure 2 As shown, Figure 1 Flowchart of a fast heterogeneous information network representation learning method based on hierarchical structure provided by an embodiment of the present invention, Figure 2 A framework diagram of a hierarchical structure-based fast heterogeneous information network representation learning method provided by an embodiment of the present invention. The method includes the following steps:

[0039] S1. Obtain the data to be mined, clean and process the data, encode the data, and build a heterogeneous information network.

[0040] In a specific implementation, step S1 specifically includes the following steps:

[0041] S11. Processing academic paper database data, specifically including the following steps:

[0042] Step 1: Extract the paper titles, authors, and citations from the academic paper database and number them sequentially. Specifically, extract 127,623 papers, 164,473 authors, and 147,251 citations from the academic paper database and number them sequentially: paper number (0, 1, ..., 127622), author number (127623, 127624, ..., 292095), and citation number (292096, 292097, ..., 439346).

[0043] Step 2: Divide the papers into several categories based on their research fields, then number the categories and set the category numbers as the paper labels. Number the paper data extracted in step 1 according to 10 categories: Comparative graphics multimedia is numbered 0, database data mining information retrieval is numbered 1, artificial intelligence is numbered 2, human computer interaction ubiquitous computing is numbered 3, interdisciplinary studies is numbered 4, theoretical computer science is numbered 5, computer networks is numbered 6, information security is numbered 7, software engineering is numbered 8, and high performance computing is numbered 9.

[0044] S12. Set papers, authors, and citations as different types of nodes. Set paper-author edges based on the relationship between authors publishing papers. Set paper-citation edges based on the citation relationship between papers. Constructed edges: 355,072 paper-author edges and 392,519 paper-citation edges. Construct a heterogeneous information network based on different types of nodes and the relationships between nodes. ,in and Represent the node set and edge set in the network respectively. For each node There is a node type mapping function For each edge There is an edge type mapping function ,in and represent the types of nodes and edges in the network, respectively, and and Specifically, a heterogeneous information network was constructed, consisting of 439,347 nodes of three types and 747,591 edges of two types. The heterogeneous information network was written into a text file using Python code, organized in the format of "node number\tnode number\n."

[0045] S2, network-based The heterogeneity and topological structure in the network can reduce the network size to obtain a small-scale heterogeneous information network. as well as and Node correspondence between The process of iteratively shrinking the network obtains a series of hierarchical heterogeneous information networks with decreasing network size. and the set of mapping matrices between network layers ,like Representation Network and The mapping matrix between Indicates the maximum number of layers in a hierarchical heterogeneous information network. In a specific implementation, step S2 specifically includes the following steps:

[0046] S21. Design a model to reduce the size of heterogeneous information networks , iterative use Reduce the network scale and obtain a series of hierarchical heterogeneous information networks with decreasing network scales For any layer of heterogeneous information network ,use Based on network heterogeneity and network topology, heterogeneous information network Small-scale heterogeneous information networks formed by reducing scale ,Right now .Model The first step is to divide the heterogeneous information network into The division result is A collection of societies ,gather Any community in It includes multiple types of nodes. In the specific operation, the community partitioning algorithm Louvain here can also use other non-overlapping community partitioning methods.

[0047] S22, Model The second step is to pass the collection Any community in China Decompose into multiple type communities according to node type, represented as ,in, Indicates the Layer 1 heterogeneous information network In the community Type community of type, a type community means that all nodes in the community are of the same type. All communities in the network are rearranged according to the node type to obtain a heterogeneous information network. The results of community division by type ,in The type community set is represented as , that is, the node types contained in the community are Type, where express There are indivual Type communities. For example, using the Louvain algorithm on an academic paper network, we can obtain multiple communities. One community contains nodes of three types: papers, authors, and citations. This community is then decomposed into three type communities: paper communities, author communities, and citation communities. The multiple communities obtained by the Louvain algorithm are then decomposed and rearranged, placing all paper communities first, followed by all author communities, and finally citation communities. This results in a variety of type communities, which are then rearranged.

[0048] S23. Based on heterogeneous information network The results of community division by type , reduce the size of the network to obtain the network .Will The type community in The nodes in the type community are aggregated into The edges between the midpoints. Based on the above idea, we can get and The mapping matrix between , The number of rows is The number of nodes in The number of columns is The number of nodes, that is The number of medium-type communities. The value of is {0,1}, for the type community All nodes in the matrix The corresponding rows in The corresponding column coordinates are treated as 1. The third step is based on and The mapping matrix between Zoom out get The specific formula is as follows:

[0049]

[0050] in, Representing heterogeneous information networks The adjacency matrix of Represents a heterogeneous information network after downscaling The adjacency matrix of .

[0051] S24, Model-based Heterogeneous information network under construction Iteratively shrink the network to obtain a series of hierarchical heterogeneous information networks with scales from large to small The obtained hierarchical network expresses the structural information in the network from different granularities, and there are associations between levels.

[0052] S3. Use representation learning algorithm to obtain the smallest Layered Heterogeneous Information Network Node representation . Design a heterogeneous graph convolutional model and based on Train the model to obtain model parameters.

[0053] In a specific implementation, step S3 specifically includes the following steps:

[0054] S31. In the smallest heterogeneous information network The existing unsupervised heterogeneous information network representation learning method is used to learn the representation of nodes ,in , Represents the dimension of the low-dimensional vector. In the specific operation, Deepwalk, metapath2vec, RHINE, HeteSpaceyWalk, MetaGraph2vec and NSHE were used to embed the coarsest layer network. It is very flexible and can use any unsupervised heterogeneous information network representation learning method embedded in all nodes in the network. Since the network size is greatly reduced by S2, the representation learning time of this step is much shorter than that of the constructed heterogeneous information network. The top indicates the time of study.

[0055] S32. Design a heterogeneous graph convolutional model, called a heterogeneous refinement model , first use the Layered Heterogeneous Information Network Node representation To train a heterogeneous refinement model , the first step is to process In order to solve the heterogeneity in the model, different types of node representations are embedded into the same space through the type parameter matrix, and then the Layer convolutional neural network is used to propagate information. No. Taking the layer neural network as an example, is a node In the neural network The hidden layer state in the layer, the specific formula is as follows:

[0056]

[0057] in, , It is The trainable weight matrix of the layer neural network, Indicates the The dimension of the layer representation, Indicates the Layer neural network with node type The corresponding type parameter matrix, Representation node of The training optimization function formula is as follows:

[0058]

[0059] in, Representing heterogeneous information networks The adjacency matrix of express The node vector of the final output of the layer convolutional neural network.

[0060] S4. Using the mapping matrix The network Node representation Pass back to the network ,get The initialization vector of the node is then obtained through the heterogeneous graph convolution model trained in S3 to obtain the final representation of the node in this layer, and the above process is repeated until the original heterogeneous information network , and finally obtain the node representation of the constructed heterogeneous information network.

[0061] In a specific implementation, step S4 specifically includes the following steps:

[0062] S41, heterogeneous refinement model trained based on S3 , the smallest heterogeneous information network A heterogeneous information network constructed layer by layer First, use the mapping relationship between different layers of the network The embedding vector of the small-scale network is mapped to the larger-scale network as the initialization node vector of the larger-scale network. Refinement For example, the specific formula is as follows:

[0063]

[0064] S42. Obtaining a heterogeneous information network Initialization vector Then, using the heterogeneous refinement model The structural information within the same layer and the heterogeneous information in the network are retained and the information between layers is transmitted during the refinement process. Refinement For example, the specific formula is as follows

[0065]

[0066] S43, based on the two steps S41 and S42, from the coarsest layer of heterogeneous information network A heterogeneous information network constructed layer by layer , S41 and S42 are executed at each layer until the final original heterogeneous information network node representation is obtained.

[0067] To verify the effectiveness and speed of the technical solution proposed in this invention, we conducted extensive experiments on three real-world datasets: ACM, AMiner_S, and Aminer_L. The details of the three datasets are provided in Table 1. In these experiments, we compared our method with existing advanced representation learning methods to demonstrate its effectiveness. The algorithms selected for comparison include DeepWalk (DW), HERec, RHINE, Metapath2vec (MP2vec), MetaGraph2vec (MG2vec), HeteSpaceyWalk (HSWalk), HAN, NSHE, HeCo, and HINChip. It should be noted that the NSHE experiment on the AMiner_L dataset, conducted on our device, took over six days. The experiments on the AMiner dataset, conducted by HAN and HINChip, both experienced memory overflow. Furthermore, due to the lack of attribute information on AMiner, the HeCo experiment could not be performed. Therefore, the corresponding experimental results are not shown in Tables 2-5.

[0068] Table 1 Dataset

[0069]

[0070] Table 2 Node classification experiments on the ACM dataset

[0071]

[0072] Table 3 Node classification experiments on the Aminer_S dataset

[0073]

[0074] Table 4 Node classification experiments on the Aminer_L dataset

[0075]

[0076] Table 5 Comparison of experimental time on three datasets

[0077]

[0078] As can be seen from Tables 2-4, the HEHE method of the present invention achieves essentially optimal results, demonstrating the effectiveness of the present method in improving the embedding accuracy of representation learning in heterogeneous information networks by constructing a heterogeneous multi-granularity network. Table 5 shows that compared to other methods, the present method is 1.44 to 54.88 times faster. This time comparison demonstrates the rapidity of the present method.

[0079] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of any of the above methods.

[0080] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any of the above methods.

[0081] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute the steps of any one of the methods in the above embodiments.

[0082] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above method.

[0083] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0084] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A hierarchical structure-based fast heterogeneous information network representation learning method, characterized by: The following steps are included: S1. Obtain the data to be mined, clean and process the data, encode the data, and build a heterogeneous information network ; Said S1 establishes a heterogeneous information network through the following steps: S11. First, extract the paper title, author and citation from the academic paper database; number them in sequence, including the paper number from 0 to , author number from arrive , citation number from arrive ; Divide the papers into several categories according to their research fields, number them, and set the category numbers as labels for the papers; S12. Set papers, authors, and citations as different types of nodes, set paper-author edges based on the relationship between authors publishing papers, and set paper-citation edges based on the relationship between papers and citations; construct a heterogeneous information network based on the aforementioned different types of nodes and the relationships between nodes ,in and Represent the node set and edge set in the network respectively, Represents a set of node and edge types, and Represents the node type mapping function and edge type mapping function respectively; for each node There is a node type mapping function ; For each edge There is an edge type mapping function ,in and represent the types of nodes and edges in the network, respectively, and and ; S2, network-based The heterogeneity and topological structure in the network can reduce the network size to obtain a small-scale heterogeneous information network. as well as and Node correspondence between ; The process of iteratively shrinking the network obtains a series of hierarchical heterogeneous information networks with decreasing network size and the set of mapping matrices between network layers ,include Representation Network and The mapping matrix between Represents the maximum number of layers in a hierarchical heterogeneous information network; S3. Use representation learning algorithm to obtain the smallest Layered Heterogeneous Information Network Node representation , design a heterogeneous graph convolutional model and based on Training the model to obtain model parameters; step S3 specifically includes: S31. In the smallest heterogeneous information network The existing unsupervised heterogeneous information network representation learning method is used to learn the representation of nodes ,in , Represents the dimension of the low-dimensional vector; S32. Design a heterogeneous graph convolutional model, called a heterogeneous refinement model , first use the Layered Heterogeneous Information Network Node representation To train a heterogeneous refinement model , the first step is to process In order to solve the heterogeneity in the model, different types of node representations are embedded into the same space through the type parameter matrix, and then the Layer convolutional neural network for information propagation; No. Taking the layer neural network as an example, is a node In the neural network The hidden layer state in the layer, the specific formula is as follows: in, , Indicates the Layer neural network with node type The corresponding type parameter matrix, Representation node of The neighbor nodes of the type; the optimization function formula for training is as follows: in, Representing heterogeneous information networks The adjacency matrix of express The node vector of the final output of the layer convolutional neural network; S4. Using the mapping matrix The network Node representation Pass back to the network ,get The initialization vector of the node is then obtained through the heterogeneous graph convolution model trained in S3 to obtain the final representation of the node in this layer, and the above process is repeated until the original heterogeneous information network , and finally obtain the node representation of the constructed heterogeneous information network.

2. The hierarchical structure-based fast heterogeneous information network representation learning method according to claim 1 is characterized by: The step S2 specifically includes: S21. Design a model to reduce the size of heterogeneous information networks , iterative use Reduce the network scale and obtain a series of hierarchical heterogeneous information networks with decreasing network scales ; For any layer of heterogeneous information network ,use Based on network heterogeneity and network topology, heterogeneous information networks Small-scale heterogeneous information networks formed by reducing scale ,Right now ;Model The first step is to divide the heterogeneous information network into The division result is A collection of societies ,gather Any community in Contains multiple types of nodes; S22, Model The second step is to pass the collection Any community in China Decompose into multiple type communities according to node type, represented as ,in, Indicates the Layer 1 heterogeneous information network In the community Type community of type, a type community means that all nodes in the community are of the same type; decomposition All communities in the network are rearranged according to the node type to obtain a heterogeneous information network. The results of community division by type ,in The type community set is represented as , that is, the node types contained in the community are Type, where express There are indivual Type Society; S23. Based on heterogeneous information network The results of community division by type , reduce the size of the network to obtain the network ;Will The type community in The nodes in the type community are aggregated into Edges between midpoints; get and The mapping matrix between , The number of rows is The number of nodes in The number of columns is The number of nodes, that is The number of medium-type communities; matrix The value of is {0,1}, for the type community All nodes in the matrix The corresponding rows in The coordinate of the corresponding column is treated as 1; The third step is based on and The mapping matrix between Zoom out get The specific formula is as follows: in, Representing heterogeneous information networks The adjacency matrix of Represents a heterogeneous information network after downscaling The adjacency matrix of S24, Model-based Heterogeneous information network under construction Iteratively shrink the network to obtain a series of hierarchical heterogeneous information networks with scales from large to small .

3. A computer device storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 2.

Images